Call analysis

Post-call analysis reads a finished call’s transcript (speech, tool calls and results) in one model call and writes a summary, an outcome judged against your definition of success, and the fields you ask it to extract. Configure it under config.analysis; it is on by default with the summary on. Chats are analysed the same way when they end (end_call or POST /v1/chats/{chat_id}/end); a chat nobody ends is never analysed.

Configure it

{
  "analysis": {
    "enabled": true,
    "summary": true,
    "outcome": {
      "prompt": "The caller's appointment was booked and they were told the date and time."
    },
    "fields": [
      {
        "name": "appointment_date",
        "type": "string",
        "description": "The date booked, as YYYY-MM-DD. Null if no appointment was made."
      },
      {
        "name": "callback_requested",
        "type": "boolean",
        "description": "True only if the caller explicitly asked to be called back."
      }
    ]
  }
}

Send it in PATCH /v1/agents/{agent_id} (fields is a list, so send it whole), then publish: analysis runs from the published version.

Field	Default	Meaning
`enabled`	`true`	Analyse sessions after they end.
`summary`	`true`	Two or three sentences: why they called, what the agent did, how it ended. Shown on the calls list.
`outcome.prompt`	unset	What a successful call looks like for this agent. Unset means calls are not judged.
`fields`	`[]`	Up to 25 values to extract.
`model`	the agent’s `llm`	An LLM `{provider, model, reasoning_effort}` to use instead. Requires that provider’s key. `priority` is refused. A realtime agent with no `model` uses the first catalog LLM from its own provider.

Each entry in fields:

Key	Rules
`name`	Required. `^[a-z][a-z0-9_]{0,39}$`. Unique. `summary`, `outcome` and `outcome_rationale` are reserved.
`type`	`string` (default), `boolean`, `integer` or `number`. No objects or arrays.
`description`	Required. The only instruction the model gets for this field: say what it is, its format, and what to return when the call never mentions it.

Every field is nullable: a value the call never established, or one returned in the wrong type, comes back null.

Cost and timing

Runs once per session, after the caller has hung up, on your provider key. Budget about 5% on top of the call’s provider cost; it is billed in the same pass as the call.
Hard 30-second timeout. A failure never affects the call.

Read the result

session.analysis on GET /v1/calls/{session_id} (analysis on GET /v1/chats/{chat_id}), and the same object in the session.completed webhook.

Field	Meaning
`status`	`none`, `pending`, `completed`, `failed` or `skipped`.
`skip_reason`	Set only with `skipped`.
`summary`	The summary, when on.
`outcome`	`success`, `failure` or `unknown`; null when no `outcome` is configured. `unknown` means the call ended before anything could be judged.
`outcome_rationale`	One or two sentences citing what decided the verdict. Read it to check your `outcome.prompt` says what you meant.
`fields`	Extracted values by name; `{}` when none configured.

Compute success rates over judged calls only: GET /v1/calls/stats returns judged_calls and successful_calls.

`status`	Meaning	Do
`none`	`enabled: false`; nothing ran or was charged.	Enable and publish.
`pending`	Running. On a call that ended a while ago, the process died mid-analysis.	Backfill it.
`completed`	Done.	—
`failed`	The model was unreachable, timed out, or returned non-JSON.	Backfill it; if it repeats, check the key for the analysis model.
`skipped`	Refused before spending anything; see `skip_reason`.	—

`skip_reason`	Why
`too_short`	Fewer than two caller turns (wrong number, hangup on the greeting, voicemail). A turn is speech or an image.
`call_failed`	The call ended on a failure close reason.
`produces_nothing`	Summary off, no outcome, no fields. Publishing that is refused; it only appears on calls under an older definition.

Preview a definition

POST /v1/calls/{session_id}/analysis/preview runs an analysis definition against one finished call and returns {analysis, provider, model}. It writes nothing, but it is a real model call on your key. It runs even when the definition has enabled: false. Pass agent_id to judge against another agent’s prompt; omit it to use the prompt the call ran. Chats: POST /v1/chats/{chat_id}/analysis/preview.

curl -X POST https://api.talqing.com/v1/calls/9c1f6ad2-3b47-4e58-9a2c-7f0d51e8c4b1/analysis/preview \
  -H "Authorization: Bearer $TALQING_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "analysis": {
      "outcome": { "prompt": "An appointment was booked and confirmed aloud." },
      "fields": [
        { "name": "party_size", "type": "integer", "description": "How many people the booking is for. Null if never stated." }
      ]
    }
  }'

const preview = await talqing.calls.analysis.preview({
  session_id: "9c1f6ad2-3b47-4e58-9a2c-7f0d51e8c4b1",
  previewAnalysisRequest: {
    analysis: {
      outcome: { prompt: "An appointment was booked and confirmed aloud." },
      fields: [
        { name: "party_size", type: "integer", description: "How many people the booking is for. Null if never stated." },
      ],
    },
  },
});
console.log(preview.model, preview.analysis.outcome, preview.analysis.fields);

preview = talqing.calls.analysis.preview(
    "9c1f6ad2-3b47-4e58-9a2c-7f0d51e8c4b1",
    analysis={
        "outcome": {"prompt": "An appointment was booked and confirmed aloud."},
        "fields": [
            {"name": "party_size", "type": "integer", "description": "How many people the booking is for. Null if never stated."}
        ],
    },
)
print(preview["model"], preview["analysis"]["outcome"], preview["analysis"]["fields"])

Re-run on past calls

POST /v1/calls/analysis/backfill with {"call_ids": [...]} (1 to 100) re-analyses each call with its agent’s saved definition, one at a time, and returns {analysed, skipped, failed}. It is also the retry path for failed and stuck pending. A call still running is a 409; a chat or unknown id is a 404. For one chat, use POST /v1/chats/{chat_id}/analysis.

Each re-run is another model call on your provider key and is added to the call’s provider_cost. The platform fee does not change. Preview first.

Analysis is not userdata

session.userdata is what the agent’s tools established during the call; analysis.fields is what a model inferred afterwards. They are never merged, so a guess cannot overwrite a confirmed value. See userdata.

Teams, resumed calls and memory

Analysis uses the configuration of the agent that answered. The transcript it reads marks every handoff ([Handed off to Billing], [Started …], [Back to …]).
A resumed web call is analysed as one call: the newest call reads every call’s transcript in the chain. Earlier calls keep their own analysis.
conversation.context: "summary" (conversation memory) feeds these summaries into later calls, so publishing it with analysis or its summary off is refused.

Publish refusals

Error	Cause
`call analysis is on but nothing is switched on to produce - …`	`enabled: true` with no summary, outcome or fields.
`call analysis has no model to run on - …`	No `llm`, no realtime provider LLM, and no `analysis.model`.
`the call analysis model cannot use the priority lane - …`	`priority: true` on `analysis.model`.
`'include summaries of past conversations' needs call analysis to be writing summaries - …`	`conversation.context: "summary"` with analysis or summary off.

In the dashboard

The call page’s Call analysis panel shows the outcome badge, summary, rationale and fields, or why there is none. Its reload control offers Re-run (saved definition, writes the result) and Try current draft (preview, writes nothing); both are hidden on a running or erased call.

Calls

The record analysis is written onto.

Webhooks

Push results to your systems.

Get started

Build

Guides

Configure it

Cost and timing

Read the result

Preview a definition

Re-run on past calls

Analysis is not userdata

Teams, resumed calls and memory

Publish refusals

In the dashboard

Calls

Webhooks

​Configure it

​Cost and timing

​Read the result

​Preview a definition

​Re-run on past calls

​Analysis is not userdata

​Teams, resumed calls and memory

​Publish refusals

​In the dashboard

Calls

Webhooks

Configure it

Cost and timing

Read the result

Preview a definition

Re-run on past calls

Analysis is not userdata

Teams, resumed calls and memory

Publish refusals

In the dashboard